metacognitive-behavioral-tuning/Qwen3-0.6B-MBT-R

TEXT GENERATIONConcurrent Unit Cost:1Model Size:0.8BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 3, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Qwen3-0.6B-MBT-R is a 0.8 billion parameter causal language model developed by metacognitive-behavioral-tuning, based on the Qwen3-0.6B architecture. This model is specifically fine-tuned using Metacognitive Behavioral Tuning (MBT-R) for multi-hop question answering, employing a refinement method that rewrites reasoning traces into a 5-phase structure for SFT and subsequent GRPO. It features a 32768 token context length and is optimized for complex reasoning tasks, particularly excelling in question answering benchmarks like HotpotQA.

Loading preview...

Overview

Qwen3-0.6B-MBT-R is a 0.8 billion parameter language model derived from the Qwen/Qwen3-0.6B base model. It is specifically fine-tuned using the Metacognitive Behavioral Tuning (MBT-R) method, which involves a two-stage process of Supervised Fine-Tuning (SFT) followed by Guided Reinforcement Learning with Policy Optimization (GRPO). The core innovation of MBT-R is its approach to refining the student model's own reasoning traces, restructuring them into a standardized 5-phase format for more effective training.

Key Capabilities

  • Multi-Hop Question Answering: Specialized for complex question answering tasks that require synthesizing information from multiple sources or steps.
  • Metacognitive Behavioral Tuning: Utilizes a unique training methodology that enhances the model's reasoning capabilities by explicitly structuring and refining its internal thought processes.
  • Performance on HotpotQA: Demonstrates strong performance on the HotpotQA benchmark (in-distribution) and shows generalization to out-of-distribution datasets like MuSiQue and 2WikiMultiHopQA.

Good For

  • Research in Reasoning: Ideal for researchers exploring advanced reasoning techniques and metacognitive processes in LLMs.
  • Complex QA Systems: Suitable for integration into applications requiring robust multi-hop question answering capabilities.
  • Educational Tools: Potentially useful for developing AI assistants that can explain reasoning steps in a structured manner.

Finetunes
1 models